This tutorial implements a Convolutional Neural Network classifier on AMD Versal™ AI Edge Series Gen 2 for radio signal classification.
The model architecture follows what is defined in [1]. DeepSig Dataset 2018.01A [2] is used to train the model. This tutorial example illustrates a number of key topics fundamental to custom coding machine learning designs using the AIE API including:
Using multi-node matrix multiply intrinsics to vectorize ConvNet layer compute workloads
Using 2D addressing patterns of memory tiles to access layer I/O in the order required for consumption by the compute
Using zero-padding capability of the memory tiles when feeding input data to ConvNet layers to preserve the original input size
Using async RTPs to send network weights from host at startup and using local tile memory for storing them
Using custom coding to implement SeLU activation function